Land cover and flood type govern the detection limits of satellite-based flood mapping across diverse global flood events

Researchers deployed the Prithvi-EO-2.0 geospatial foundation model across 19 diverse flood events globally to assess satellite-based flood detection reliability. The study found that detection accuracy varies significantly by land cover type and flood mechanism, with cropland showing the highest accuracy (IoU=52%) while tree cover and built-up areas achieved near-zero detection (IoU=4%), establishing critical operational boundaries for disaster response systems.

This research addresses a critical gap in understanding the real-world performance of AI-powered satellite flood mapping systems deployed for emergency response. The Prithvi-EO-2.0 model, a geospatial foundation model pretrained on extensive satellite archives, demonstrates the promise and limitations of geographic transferability in machine learning—a key concern as AI systems scale to handle diverse, unseen environmental conditions. The study's validation across 19 out-of-distribution events spanning multiple continents and climate zones provides unprecedented empirical grounding for operational reliability claims.

The findings reveal environment-dependent detection boundaries that fundamentally reshape expectations for AI-assisted disaster response. Cropland environments enable reliable flood detection while tree-covered and urban areas present near-insurmountable challenges, suggesting that model capacity alone cannot overcome physical constraints in satellite imagery interpretation. The dual-reference validation methodology proves particularly valuable, showing that apparent model failures sometimes reflect inconsistencies between reference datasets rather than actual detection failures—a distinction with significant implications for model evaluation practices.

For organizations deploying satellite-based flood monitoring systems, these results establish realistic performance expectations across different geographic contexts. The identification of 23 failure modes, with pipeline engineering errors exceeding model limitations, highlights that operational challenges often lie outside the neural network itself. This has direct implications for disaster response agencies planning flood mapping deployments, suggesting that system architecture and data preprocessing merit equal attention to model selection. The research provides a framework for assessing when satellite-based detection becomes unreliable, enabling more honest risk assessment in climate adaptation planning.

• →Satellite flood detection accuracy depends jointly on land cover type and flood mechanism, not uniformly across environments.
• →Cropland shows highest detection reliability (IoU=52%) while tree cover and urban areas achieve near-zero detection (IoU=4%).
• →Dual-reference validation revealed that some apparent model errors stem from definitional inconsistencies between reference products rather than detection failures.
• →Pipeline engineering failures dominated initial error sources over model capacity limitations in the tested system.
• →The study establishes operational detection boundaries critical for realistic expectations in disaster response planning.